Google administrators have scrambled to correct AI errors that have seen users told to eat rocks and mix pizza cheese with glue.
The California-based tech giant has come under fire after its proprietary AI, AI Overview, came up with bizarre answers to questions posed by people using the service.
The engine’s newest feature is designed to make searching for information more straightforward by collating entries from across the web and presenting them on the top of the first results page.
But some of these initial responses have pulled information from sites that only vaguely resemble the content of the questions.
They have ranged from potentially toxic cooking suggestions to wildly incorrect historical facts.
In one viral piece of advice from the AI, users were advised to mix glue with cheese to make it stick to their pizza.
Eating glue, even the non-toxic kind, is dangerous, and the advice appears to have originated in a joke post on the social media platform Reddit in 2013.
The AI response even suggested people add an eighth of a cup to the pizza sauce to “give it more tackiness”.
Another response saw the tool claim that geologists recommend humans consume at least one rock per day.
And in one non-nutrition-based search response, the AI claimed that 17 of the 42 US presidents were white.
One user posting on X, formerly Twitter, shared a picture of a result which claimed that doctors recommend smoking between two and three cigarettes per day during pregnancy.
In reality, smoking while pregnant has been long deemed dangerous to unborn children and expectant mothers, who are placed at an increased risk of pregnancy-related problems like miscarriage and premature labour.
Google has said the viral answers are not representative of AI Overview’s overall performance, as the examples are “very uncommon queries”, meaning they “aren’t representative of most people’s experiences”.
A spokesperson for the firm said: “The vast majority of AI Overviews provide high-quality information, with links to dig deeper on the web. We conducted extensive testing before launching this new experience to ensure AI overviews meet our high bar for quality.
“Where there have been violations of our policies, we’ve taken action – and we’re also using these isolated examples as we continue to refine our systems overall.”
